For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo package

sm100

SM100 Structured Convolution Kernels.

High-performance Conv2D for NVIDIA Blackwell (SM100) GPUs using implicit GEMM with hardware TMA im2col. Reuses infrastructure from sm100_structured matmul.

Supported: Conv2D fprop with stride=1, dilation=1, BF16/FP16.

Modules

• ​conv2d: Public API for SM100 Conv2D forward propagation.
• ​conv2d_fprop_kernel: SM100 Conv2D Forward Propagation Kernel.
• ​conv_config: Configuration types for SM100 structured convolution kernels.
• ​conv_smem: Shared memory layout for SM100 Conv2D kernel.
• ​conv_tile_loader: Tile loader for SM100 convolution with hardware im2col TMA.
• ​dispatch: SM100 Conv2D dispatch for the nn conv op.
• ​epilogue_load_pipeline: Epilogue load pipeline types for SM100 Conv2D kernel.
• ​qslice_conv3d: Q-slice 3-D conv → Q sequential SM100 2-D conv calls with fp32 accumulator.